摘要： The dynamic characteristic of an autonomous underwater vehicle (AUV) is affected when it is reconfigured with different payloads. It is desirable to have an updated model, such that the control and guidance law can be redesigned to obtain better performance. Hence, we develop a method to enable online identification of AUV dynamics via in-field experiments, where the AUV is commanded to execute a compact set of maneuvers under doublet excitation. The identification process has two stages. In the training stage, state variable filter and recursive least square (SVF-RLS) estimator is used to estimate the unknown parameters. In the validation stage, the prediction capability of the model is checked using a fresh data set. The parameters converged within 12 s in the experiments using five different thrusts. Validation results show that the identified models are able to explain 78% to 92% of the output variation. Next, we compare the SVF-RLS estimator with the conventional offline identification method. The comparison shows that the SVF-RLS estimator is better in terms of prediction accuracy, computational cost and training time. The usefulness of the identified models is highlighted in two applications. We use it to estimate the turning radius of the AUV at different speeds, and to design a gain-scheduled controller.